Spaces:
Runtime error
Runtime error
replace sox with librosa
Browse files- app.py +14 -16
- requirements.txt +6 -4
app.py
CHANGED
@@ -1,23 +1,24 @@
|
|
1 |
import soundfile as sf
|
|
|
2 |
import torch
|
3 |
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
|
4 |
import gradio as gr
|
5 |
-
import sox
|
6 |
import os
|
7 |
|
8 |
|
9 |
-
def convert(inputfile, outfile):
|
10 |
-
sox_tfm = sox.Transformer()
|
11 |
-
sox_tfm.set_output_format(
|
12 |
-
file_type="wav", channels=1, encoding="signed-integer", rate=16000, bits=16
|
13 |
-
)
|
14 |
-
sox_tfm.build(inputfile, outfile)
|
15 |
-
|
16 |
-
|
17 |
api_token = os.getenv("API_TOKEN")
|
18 |
model_name = "indonesian-nlp/wav2vec2-indonesian-javanese-sundanese"
|
19 |
processor = Wav2Vec2Processor.from_pretrained(model_name, use_auth_token=api_token)
|
20 |
model = Wav2Vec2ForCTC.from_pretrained(model_name, use_auth_token=api_token)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
|
22 |
|
23 |
def parse_transcription(wav_file):
|
@@ -25,6 +26,7 @@ def parse_transcription(wav_file):
|
|
25 |
convert(wav_file.name, filename + "16k.wav")
|
26 |
speech, _ = sf.read(filename + "16k.wav")
|
27 |
input_values = processor(speech, sampling_rate=16_000, return_tensors="pt").input_values
|
|
|
28 |
logits = model(input_values).logits
|
29 |
predicted_ids = torch.argmax(logits, dim=-1)
|
30 |
transcription = processor.decode(predicted_ids[0], skip_special_tokens=True)
|
@@ -37,14 +39,10 @@ input_ = gr.inputs.Audio(source="microphone", type="file")
|
|
37 |
|
38 |
gr.Interface(parse_transcription, inputs=input_, outputs=[output],
|
39 |
analytics_enabled=False,
|
40 |
-
show_tips=False,
|
41 |
-
theme='huggingface',
|
42 |
-
layout='vertical',
|
43 |
title="Multilingual Speech Recognition for Indonesian Languages",
|
44 |
description="Automatic Speech Recognition Live Demo for Indonesian, Javanese and Sundanese Language",
|
45 |
-
article="
|
46 |
-
"This demo was built for the project "
|
47 |
"<a href='https://github.com/indonesian-nlp/multilingual-asr' target='_blank'>Multilingual Speech Recognition for Indonesian Languages</a>. "
|
48 |
"It uses the <a href='https://huggingface.co/indonesian-nlp/wav2vec2-indonesian-javanese-sundanese' target='_blank'>indonesian-nlp/wav2vec2-indonesian-javanese-sundanese</a> model "
|
49 |
-
"which was fine-tuned on Indonesian Common Voice, Javanese and Sundanese OpenSLR speech datasets."
|
50 |
-
|
|
|
1 |
import soundfile as sf
|
2 |
+
import librosa
|
3 |
import torch
|
4 |
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
|
5 |
import gradio as gr
|
|
|
6 |
import os
|
7 |
|
8 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
9 |
api_token = os.getenv("API_TOKEN")
|
10 |
model_name = "indonesian-nlp/wav2vec2-indonesian-javanese-sundanese"
|
11 |
processor = Wav2Vec2Processor.from_pretrained(model_name, use_auth_token=api_token)
|
12 |
model = Wav2Vec2ForCTC.from_pretrained(model_name, use_auth_token=api_token)
|
13 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
14 |
+
model = model.to(device)
|
15 |
+
|
16 |
+
|
17 |
+
def convert(inputfile, outfile):
|
18 |
+
target_sr = 16000
|
19 |
+
data, sample_rate = librosa.load(inputfile)
|
20 |
+
data = librosa.resample(data, orig_sr=sample_rate, target_sr=target_sr)
|
21 |
+
sf.write(outfile, data, target_sr)
|
22 |
|
23 |
|
24 |
def parse_transcription(wav_file):
|
|
|
26 |
convert(wav_file.name, filename + "16k.wav")
|
27 |
speech, _ = sf.read(filename + "16k.wav")
|
28 |
input_values = processor(speech, sampling_rate=16_000, return_tensors="pt").input_values
|
29 |
+
input_values = input_values.to(device)
|
30 |
logits = model(input_values).logits
|
31 |
predicted_ids = torch.argmax(logits, dim=-1)
|
32 |
transcription = processor.decode(predicted_ids[0], skip_special_tokens=True)
|
|
|
39 |
|
40 |
gr.Interface(parse_transcription, inputs=input_, outputs=[output],
|
41 |
analytics_enabled=False,
|
|
|
|
|
|
|
42 |
title="Multilingual Speech Recognition for Indonesian Languages",
|
43 |
description="Automatic Speech Recognition Live Demo for Indonesian, Javanese and Sundanese Language",
|
44 |
+
article="This demo was built for the project "
|
|
|
45 |
"<a href='https://github.com/indonesian-nlp/multilingual-asr' target='_blank'>Multilingual Speech Recognition for Indonesian Languages</a>. "
|
46 |
"It uses the <a href='https://huggingface.co/indonesian-nlp/wav2vec2-indonesian-javanese-sundanese' target='_blank'>indonesian-nlp/wav2vec2-indonesian-javanese-sundanese</a> model "
|
47 |
+
"which was fine-tuned on Indonesian Common Voice, Javanese and Sundanese OpenSLR speech datasets."
|
48 |
+
).launch( inline=False, server_name="0.0.0.0", show_tips=False, enable_queue=True)
|
requirements.txt
CHANGED
@@ -1,6 +1,8 @@
|
|
1 |
-
gradio==2.4.1
|
2 |
-
soundfile
|
3 |
torch
|
|
|
4 |
transformers
|
5 |
-
|
6 |
-
|
|
|
|
|
|
|
|
|
|
|
|
1 |
torch
|
2 |
+
tokenizers
|
3 |
transformers
|
4 |
+
datasets
|
5 |
+
gradio>=3.0.0
|
6 |
+
soundfile
|
7 |
+
sentencepiece
|
8 |
+
librosa
|